Extreme CUDA processing?

Gregory Pharr
Gregory Pharr
Joined: 15 Feb 06
Posts: 3
Credit: 126968
RAC: 0
Topic 194144

Hi. I am building a new pc that will have 3 nVidia GTX 295's running in SLI. Each board has 480 cores in parallel and almost 2Gb of dedicated RAM. Each is capable of 1.788 Teraflops of processing power. Altogether, by running 3 in SLI, this will give my machine 1440 cores and a whopping 5.364 Teraflops of power! Yes, it's like having a supercomputer sitting on your desk. But, unless your project is able to run CUDA (which the latest versions of BOINC support now), you can't make use of any of that power. It would represent a massive increase in processing power for work units. Most of the newer nVidia graphics cards support CUDA. Any plans to add CUDA support for your project? or will I have to keep using my CUDA power for other projects? Thanks!

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2143
Credit: 2951320283
RAC: 691750
mikey
mikey
Joined: 22 Jan 05
Posts: 12663
Credit: 1839061411
RAC: 4247

RE: Hi. I am building a new

Quote:
Hi. I am building a new pc that will have 3 nVidia GTX 295's running in SLI. Each board has 480 cores in parallel and almost 2Gb of dedicated RAM. Each is capable of 1.788 Teraflops of processing power. Altogether, by running 3 in SLI, this will give my machine 1440 cores and a whopping 5.364 Teraflops of power! Yes, it's like having a supercomputer sitting on your desk. But, unless your project is able to run CUDA (which the latest versions of BOINC support now), you can't make use of any of that power. It would represent a massive increase in processing power for work units. Most of the newer nVidia graphics cards support CUDA. Any plans to add CUDA support for your project? or will I have to keep using my CUDA power for other projects? Thanks!

Rumors abound but word is they are "working on it" for Einstein. Since the powers that be have not said anything yet I would guess it is not ready quite yet.

ohiomike
ohiomike
Joined: 4 Nov 06
Posts: 80
Credit: 6453639
RAC: 0

RE: RE: Hi. I am building

Message 90133 in response to message 90132

Quote:
Quote:
Hi. I am building a new pc that will have 3 nVidia GTX 295's running in SLI. Each board has 480 cores in parallel and almost 2Gb of dedicated RAM. Each is capable of 1.788 Teraflops of processing power. Altogether, by running 3 in SLI, this will give my machine 1440 cores and a whopping 5.364 Teraflops of power! Yes, it's like having a supercomputer sitting on your desk. But, unless your project is able to run CUDA (which the latest versions of BOINC support now), you can't make use of any of that power. It would represent a massive increase in processing power for work units. Most of the newer nVidia graphics cards support CUDA. Any plans to add CUDA support for your project? or will I have to keep using my CUDA power for other projects? Thanks!

Rumors abound but word is they are "working on it" for Einstein. Since the powers that be have not said anything yet I would guess it is not ready quite yet.


My hope is that when it appears, that it will be more stable than SETI's. I have tried to run several different CPU/GPU combos on SETI and have returned a bunch of bad results before shutting them down. In my experience, Einstein does not rush to release "junk" beta code just for the sake of releasing it. For the time being, my GPU machine (twin 8800 GTS cards) is running mainly Folding@Home (which appears to be very stable.. although a pain to set up for multiple GPU's).
PS- It won't hurt to wait until they get the bugs out of Boinc either. V6.45 had bugs that kept me from running it. V6.5 runs, but has a nasty bug on an X64 Windows machine- boincmgr uses 25% of the CPU, even after you close it (the task keeps running after the GUI closes). You have to use the task manager to kill it.


mikey
mikey
Joined: 22 Jan 05
Posts: 12663
Credit: 1839061411
RAC: 4247

RE: RE: RE: Hi. I am

Message 90134 in response to message 90133

Quote:
Quote:
Quote:
Hi. I am building a new pc that will have 3 nVidia GTX 295's running in SLI. Each board has 480 cores in parallel and almost 2Gb of dedicated RAM. Each is capable of 1.788 Teraflops of processing power. Altogether, by running 3 in SLI, this will give my machine 1440 cores and a whopping 5.364 Teraflops of power! Yes, it's like having a supercomputer sitting on your desk. But, unless your project is able to run CUDA (which the latest versions of BOINC support now), you can't make use of any of that power. It would represent a massive increase in processing power for work units. Most of the newer nVidia graphics cards support CUDA. Any plans to add CUDA support for your project? or will I have to keep using my CUDA power for other projects? Thanks!

Rumors abound but word is they are "working on it" for Einstein. Since the powers that be have not said anything yet I would guess it is not ready quite yet.


My hope is that when it appears, that it will be more stable than SETI's. I have tried to run several different CPU/GPU combos on SETI and have returned a bunch of bad results before shutting them down. In my experience, Einstein does not rush to release "junk" beta code just for the sake of releasing it. For the time being, my GPU machine (twin 8800 GTS cards) is running mainly Folding@Home (which appears to be very stable.. although a pain to set up for multiple GPU's).
PS- It won't hurt to wait until they get the bugs out of Boinc either. V6.45 had bugs that kept me from running it. V6.5 runs, but has a nasty bug on an X64 Windows machine- boincmgr uses 25% of the CPU, even after you close it (the task keeps running after the GUI closes). You have to use the task manager to kill it.

I too run folding with my video cards. I have 2 cards, not sli'd, a 9600GSO and a 9800GT. I too would bring them over to Einstein if their was a way to do it while keeping the cpu's and the video cards crunching full time. I do not want a cpu crunching at only 10% just so I can crunch with a video card.

Jord
Joined: 26 Jan 05
Posts: 2952
Credit: 5893653
RAC: 167

RE: I too would bring them

Message 90135 in response to message 90134

Quote:
I too would bring them over to Einstein if their was a way to do it while keeping the cpu's and the video cards crunching full time. I do not want a cpu crunching at only 10% just so I can crunch with a video card.


The problem at the moment is that the task has to be updated in videoRAM and this doesn't happen by magic, the CPU is used for that. So the present solution is to exclude one of your CPUs to only work with the GPU. That CPU or core is only partially used, the rest of it is free.

This is done because when the task in videoRAM has to be written back to the disk, exchanged for another task, you want this to happen as quickly as possible. The GPU can only run at full whack or not at all, there's no way yet to gradually increase its use. So if the GPU has to wait for the CPU to free up enough resources that it can send the data from the disk to the videoRAM, there's a good chance the task will time out and exit with an error.

6dj72cn8
6dj72cn8
Joined: 24 Jan 06
Posts: 24
Credit: 13321065
RAC: 0

I haven't been keeping up

I haven't been keeping up with this subject so I apologise if this has been asked before. Is there an ATI equivalent of NVidia's CUDA? If so, what, where and when? If not, why not? It must be in an FAQ somewhere but the internet is . . . big. Thanks.

tullio
tullio
Joined: 22 Jan 05
Posts: 2118
Credit: 61407735
RAC: 0

RE: I haven't been keeping

Message 90137 in response to message 90136

Quote:
I haven't been keeping up with this subject so I apologise if this has been asked before. Is there an ATI equivalent of NVidia's CUDA? If so, what, where and when? If not, why not? It must be in an FAQ somewhere but the internet is . . . big. Thanks.


I believe there is something called Stream. But I would wait for OpenCL (not to be confused with OpenGL) which should work on all graphic boards.
Tullio

mikey
mikey
Joined: 22 Jan 05
Posts: 12663
Credit: 1839061411
RAC: 4247

RE: I haven't been keeping

Message 90138 in response to message 90136

Quote:
I haven't been keeping up with this subject so I apologise if this has been asked before. Is there an ATI equivalent of NVidia's CUDA? If so, what, where and when? If not, why not? It must be in an FAQ somewhere but the internet is . . . big. Thanks.

Yes there is but only, so far, at Folding@Home and only for Windows machines
http://folding.stanford.edu/English/DownloadWinOther

I do the NVidia crunching with it and it is fast AND I can use both cpus on a dual core system for a Boinc project at the same time as the video card is crunching for Folding.

mikey
mikey
Joined: 22 Jan 05
Posts: 12663
Credit: 1839061411
RAC: 4247

RE: RE: I haven't been

Message 90139 in response to message 90137

Quote:
Quote:
I haven't been keeping up with this subject so I apologise if this has been asked before. Is there an ATI equivalent of NVidia's CUDA? If so, what, where and when? If not, why not? It must be in an FAQ somewhere but the internet is . . . big. Thanks.

I believe there is something called Stream. But I would wait for OpenCL (not to be confused with OpenGL) which should work on all graphic boards.
Tullio

If I am reading this correctly you are thinking of the number of processors, called 'stream processors" on each video card. The more stream processors the faster the crunching. If you go here http://www.nvidia.com/object/cuda_learn_products.html you will see a list of CUDA NVidia cards that one can crunch with. If you click on each card and then click on the Specifications tab you will see how many processor cores that card has. There is even a link to buy a card but I would use it as a guideline only, I have seen them cheaper, and more expensive, elsewhere.

tullio
tullio
Joined: 22 Jan 05
Posts: 2118
Credit: 61407735
RAC: 0

RE: If I am reading this

Message 90140 in response to message 90139

Quote:

If I am reading this correctly you are thinking of the number of processors, called 'stream processors" on each video card. The more stream processors the faster the crunching. If you go here http://www.nvidia.com/object/cuda_learn_products.html you will see a list of CUDA NVidia cards that one can crunch with. If you click on each card and then click on the Specifications tab you will see how many processor cores that card has. There is even a link to buy a card but I would use it as a guideline only, I have seen them cheaper, and more expensive, elsewhere.


I should have written "Streaming" which is an ATI SDK.
Tullio
No. Stream SDK is the right word.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.